NG_SOCKET(4) BSD Kernel Interfaces Manual NG_SOCKET(4)
ng_socket -- netgraph socket node type
A socket node is both a BSD socket and a netgraph node. The ng_socket
node type allows user-mode processes to participate in the kernel
netgraph(4) networking subsystem using the BSD socket interface. The
process must have root privileges to be able to create netgraph sockets
however once created, any process that has one may use it.
A new ng_socket node is created by creating a new socket of type
NG_CONTROL in the protocol family PF_NETGRAPH, using the socket(2) system
call. Any control messages received by the node and not having a cookie
value of NGM_SOCKET_COOKIE are received by the process, using
recvfrom(2); the socket address argument is a struct sockaddr_ng contain-
ing the sender's netgraph address. Conversely, control messages can be
sent to any node by calling sendto(2), supplying the recipient's address
in a struct sockaddr_ng. The bind(2) system call may be used to assign a
global netgraph name to the node.
To transmit and receive netgraph data packets, a NG_DATA socket must also
be created using socket(2) and associated with a ng_socket node. NG_DATA
sockets do not automatically have nodes associated with them; they are
bound to a specific node via the connect(2) system call. The address
argument is the netgraph address of the ng_socket node already created.
Once a data socket is associated with a node, any data packets received
by the node are read using recvfrom(2) and any packets to be sent out
from the node are written using sendto(2). In the case of data sockets,
the struct sockaddr_ng contains the name of the hook on which the data
was received or should be sent.
As a special case, to allow netgraph data sockets to be used as stdin or
stdout on naive programs, a sendto(2) with a NULL sockaddr pointer, a
send(2) or a write(2) will succeed in the case where there is exactly ONE
hook attached to the socket node, (and thus the path is unambiguous).
There is a user library that simplifies using netgraph sockets; see
This node type supports hooks with arbitrary names (as long as they are
unique) and always accepts hook connection requests.
This node type supports the generic control messages, plus the following:
When the last hook is removed from this node, it will shut down as
if it had received a NGM_SHUTDOWN message. Attempts to access the
sockets associated will return ENOTCONN.
This is the default mode. When the last hook is removed, the node
will continue to exist, ready to accept new hooks until it is
explicitly shut down.
All other messages with neither the NGM_SOCKET_COOKIE or
NGM_GENERIC_COOKIE will be passed unaltered up the NG_CONTROL socket.
This node type shuts down and disappears when both the associated
NG_CONTROL and NG_DATA sockets have been closed, or a NGM_SHUTDOWN con-
trol message is received. In the latter case, attempts to write to the
still-open sockets will return ENOTCONN. If the NGM_SOCK_CMD_NOLINGER
message has been received, closure of the last hook will also initiate a
shutdown of the node.
It is not possible to reject the connection of a hook, though any data
received on that hook can certainly be ignored.
The controlling process is not notified of all events that an in-kernel
node would be notified of, e.g. a new hook, or hook removal. Some node-
initiated messages should be defined for this purpose (to be sent up the
socket(2), netgraph(3), netgraph(4), ng_ksocket(4), ngctl(8)
The ng_socket node type was implemented in FreeBSD 4.0.
Julian Elischer <julianATFreeBSD.org>
BSD January 19, 1999 BSD